Public: Glean Assistant enterprise RAG stack

Architectural notes

Per their engineering blog post on building an enterprise AI assistant,
Glean uses hybrid retrieval and reranking that combines dozens of ranking
signals (semantic similarity, keyword match, document freshness,
personalization) and deliberately chose RAG over fine-tuning LLMs on
proprietary data, which they found increases hallucinations.

The same post documents custom embedding encoders fine-tuned per customer
(outperforming generic embedders), a knowledge graph with Content, People
and Activity nodes used for personalized ranking, a permission-aware index
so the LLM never sees content a user cannot access, and 100+ connectors
feeding a central index rather than federated search. Per official docs,
admins enable models via the Model Hub (35+ models including GPT, Claude
Sonnet, Gemini, Llama) and users pick the model per conversation -- hence
model: user-configurable. The retrieval stack (in-house embedders, hybrid
index, knowledge graph) is proprietary and not catalog-mappable, so no
retrieval slug is listed; eval and observability internals are not publicly
documented.